mirrors/pytorch-image-models
1
0
Fork 0
mirror of https://github.com/huggingface/pytorch-image-models.git synced 2025-06-03 15:01:08 +08:00
Code Issues Projects Releases Wiki Activity

More test fixes, pool size for 256x256 maxvit models

Browse Source
This commit is contained in:
Ross Wightman 2022-08-23 13:38:26 -07:00
parent e939ed19b9
commit cac0a4570a
3 changed files with 9 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tests/test_models.py
View File

@ -28,7 +28,7 @@ NON_STD_FILTERS = [
'vit_*', 'tnt_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', '*mixer_*', 'gmlp_*', 'resmlp_*', 'twins_*',
'convit_*', 'levit*', 'visformer*', 'deit*', 'jx_nest_*', 'nest_*', 'xcit_*', 'crossvit_*', 'beit_*',
'poolformer_*', 'volo_*', 'sequencer2d_*', 'swinv2_*', 'pvt_v2*', 'mvitv2*', 'gcvit*', 'efficientformer*',
'coatne?t_*', 'max?vit_*',
'coatnet*', 'coatnext*', 'maxvit*', 'maxxvit*',
]
NUM_NON_STD = len(NON_STD_FILTERS)

2
timm/models/efficientformer.py
View File

@ -29,7 +29,7 @@ def _cfg(url='', **kwargs):
'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None, 'fixed_input_size': True,
'crop_pct': .95, 'interpolation': 'bicubic',
'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
'first_conv': 'stem.conv1', 'classifier': 'head',
'first_conv': 'stem.conv1', 'classifier': ('head', 'head_dist'),
**kwargs
}

13
timm/models/maxxvit.py
View File

@ -94,6 +94,7 @@ default_cfgs = {
'coatnet_rmlp_0_rw_224': _cfg(url=''),
'coatnet_rmlp_1_rw_224': _cfg(
url=''),
'coatnet_nano_cc_224': _cfg(url=''),
'coatnext_nano_rw_224': _cfg(url=''),
# Trying to be like the CoAtNet paper configs
@ -105,12 +106,12 @@ default_cfgs = {
'coatnet_5_224': _cfg(url=''),
# Experimental configs
'maxvit_pico_rw_256': _cfg(url='', input_size=(3, 256, 256)),
'maxvit_nano_rw_256': _cfg(url='', input_size=(3, 256, 256)),
'maxvit_pico_rw_256': _cfg(url='', input_size=(3, 256, 256), pool_size=(8, 8)),
'maxvit_nano_rw_256': _cfg(url='', input_size=(3, 256, 256), pool_size=(8, 8)),
'maxvit_tiny_rw_224': _cfg(url=''),
'maxvit_tiny_rw_256': _cfg(url='', input_size=(3, 256, 256)),
'maxvit_tiny_cm_256': _cfg(url='', input_size=(3, 256, 256)),
'maxxvit_nano_rw_256': _cfg(url='', input_size=(3, 256, 256)),
'maxvit_tiny_rw_256': _cfg(url='', input_size=(3, 256, 256), pool_size=(8, 8)),
'maxvit_tiny_cm_256': _cfg(url='', input_size=(3, 256, 256), pool_size=(8, 8)),
'maxxvit_nano_rw_256': _cfg(url='', input_size=(3, 256, 256), pool_size=(8, 8)),
# Trying to be like the MaxViT paper configs
'maxvit_tiny_224': _cfg(url=''),
@ -1052,7 +1053,6 @@ class PartitionAttention(nn.Module):
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def _partition_attn(self, x):
C = x.shape[-1]
img_size = x.shape[1:3]
if self.partition_block:
partitioned = window_partition(x, self.partition_size)
@ -1415,6 +1415,7 @@ class Stem(nn.Module):
self.norm1 = norm_act_layer(out_chs[0])
self.conv2 = create_conv2d(out_chs[0], out_chs[1], kernel_size, stride=1)
@torch.jit.ignore
def init_weights(self, scheme=''):
named_apply(partial(_init_conv, scheme=scheme), self)